💻 computer science

Frequency-Guided Real-Time Detection of Weak Water-Surface Targetsfor Unmanned Surface Vehicles

This paper proposes a lightweight, real-time detector for unmanned surface vehicles that enhances weak water-surface target detection by integrating frequency-guided feature routing and scale-normalized adaptive localization supervision to address challenges like low contrast and small object size while maintaining strict latency and model-size constraints.

Original authors: Ling Qin, AnChuan Wang, Qing Huang, Qun Zou

Published 2026-07-15

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Ling Qin, AnChuan Wang, Qing Huang, Qun Zou

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are the eyes of a tiny, self-driving boat (called an Unmanned Surface Vehicle, or USV) zooming across a lake. Its job is to spot floating trash, rescue buoys, or other small objects. But here's the catch: the water is a nightmare for cameras. Waves ripple, sunlight glares, and the objects are often so small they might only be a few pixels wide—like trying to find a single grain of sand in a bucket of glittering confetti.

Most computer vision systems try to solve this by just making the "brain" bigger or looking harder at everything. But this paper suggests that's like trying to fix a broken radio by turning up the volume on the whole station. Instead, the authors, Ling Qin and their team, propose a smarter approach: tune the radio to the specific frequency where the signal hides.

The Big Discovery: Water Has a "Rhythm"

The team started by doing some digital detective work. They took thousands of photos of water and compared them to photos of regular scenes (like streets or parks). They ran a special math test called a "spectral analysis" to see what kind of "energy" the images had.

They found something cool: water scenes have a unique rhythm. Unlike regular pictures where details fade away quickly, water images hold onto a lot of mid-frequency energy. Think of it like this: if a regular photo is a smooth jazz song, a water photo is a song with a persistent, choppy beat (the waves) that drowns out the quiet melody (the tiny target).

The paper explicitly argues against the idea that we should just treat waves and reflections as random "noise" to be ignored. Instead, the authors found that these waves have a specific structure. They also argue against the idea that we need to make the whole computer model huge and heavy to solve this. Their solution is lightweight, suggesting that targeted tuning works better than brute force.

The Solution: The "Frequency Detective" Module

To fix the detection problem, they built a special tool called the BSSR module (Band-Selective and Spatially Routed). Imagine this module as a super-smart filter that only gets to work on the "high-resolution" part of the image where tiny objects live.

Here is how it works, using a playful analogy:

Frequency Splitting: It takes the image and splits it into three layers: low-frequency (the big, smooth waves), mid-frequency (the choppy texture where targets hide), and high-frequency (the sharp edges).
Directional Routing: It knows that water waves often move in straight lines (horizontal or vertical). So, it uses special "Sobel" filters (think of them as tiny combs) to catch edges in four directions: up, down, left, and right. This helps it spot a tiny boat even if the waves are breaking right next to it.
Adaptive Gating: This is the smartest part. The module has a "gatekeeper" that decides how much of each layer to keep. If the water is too noisy, it tightens the gate. If a target looks promising, it opens the gate wide. It does this without slowing down the boat's computer.

The paper proves this works by showing that adding this module to a standard detector (YOLO11n) boosted the detection accuracy for small targets significantly, without making the model much bigger. The model size only grew by 0.121 M parameters, which is tiny in the world of AI.

The "Smart Coach" for Drawing Boxes

Detecting an object is one thing; drawing a perfect box around it is another. Usually, the computer uses a standard rule (called CIoU loss) to learn how to draw these boxes. But the authors realized this rule is too rigid. It's like a coach who yells the same amount of correction whether you are totally lost or just slightly off-target.

They introduced a new rule called SNK-IoU. Imagine a coach who uses a "Kalman gain" (a fancy math term for a smart adjustment knob).

If the computer is way off (low overlap), the coach turns the knob up and gives a strong correction.
If the computer is almost right (high overlap), the coach turns the knob down and gives a fine-tune adjustment.

Crucially, this "coach" only works while the computer is learning (training). Once the boat is out on the water, the coach disappears, so it doesn't slow down the real-time detection at all. The paper shows this method helped the model learn faster and spot more targets, improving the accuracy score (mAP50:95) from 41.44% to 47.35%.

How Sure Are They?

The authors are quite confident in their numbers, but they are also very honest about the limits.

What they proved: On their test datasets (FloW and PoTATO), their method consistently beat the standard models. They measured the speed on powerful desktop computers (NVIDIA RTX 5090 and 4080 Super) and found it runs at 93.1 FPS (frames per second) on the big computer and 156.2 FPS when exported to a standard format (ONNX). This suggests it is fast enough for real-time use on a desktop.
What they don't claim: They explicitly state they have not tested this on the tiny, embedded computers (like Jetson Orin) that would actually be on a real boat. They warn that the real-world speed on those small chips might be different.
What they rule out: They admit their system isn't magic. If a target is smaller than 8×8 pixels, or if the water is pitch black, foggy, or has a blinding glare that looks exactly like a target, the system might still fail. They also note that the FloW dataset they used doesn't have a separate "final exam" test set, so they had to rely on cross-testing with a different dataset (PoTATO) to prove it works generally.

The Bottom Line

This paper suggests that to find tiny, weak objects on water, we shouldn't just make our AI bigger. Instead, we should listen to the "frequency" of the water and build a detector that knows exactly where to look. By combining a frequency-guided filter (BSSR) with a smart, adaptive coach (SNK-IoU), they created a system that is faster, smaller, and better at spotting the "needles in the haystack" of the ocean.

While it's not a fully solved problem for every weather condition or every tiny computer chip yet, the results suggest that tuning into the specific "rhythm" of the water is a powerful new way to help self-driving boats see what's floating in front of them.

Problem Statement

Real-time visual perception for Unmanned Surface Vehicles (USVs) faces a critical challenge: detecting weak, small water-surface targets (e.g., floating debris, rescue objects) under low Signal-to-Noise Ratio (SNR) conditions. These targets often occupy only a few pixels (e.g., 12×12 to 32×32 pixels) and are obscured by waves, specular reflections, and non-uniform illumination. Existing detectors struggle with this "weak-signal interpretation" problem because:

Frequency Mismatch: Many detectors treat wave interference and reflections as generic visual clutter without analyzing their specific frequency structure.
Global vs. Local: Frequency-aware models often enhance features globally, whereas weak floating targets require selective enhancement specifically in the high-resolution small-target branches.
Static Supervision: Standard bounding-box regression losses (e.g., CIoU) do not adapt their correction strength based on the specific prediction quality or geometric deviation of individual samples.

Methodology

The authors propose a lightweight detector built upon the YOLO11n architecture, introducing two primary innovations: a frequency-guided feature module and an adaptive regression loss.

1. Spectral Analysis of Water-Surface Scenes

Before designing the model, the authors conducted a cross-dataset spectral analysis comparing water-surface datasets (FloW, FML-v1, FML-v2) against generic scenes (COCO).

Findings: Water-surface images exhibit significantly lower spectral decay exponents ( $\beta \approx 2.3\text{--}2.5$ ) compared to generic scenes ( $\beta \approx 2.7$ ), indicating slower spectral decay and higher energy retention in mid- and high-frequency bands.
SNR: The local SNR for small water targets is extremely low (3–8 dB) compared to generic objects (12–18 dB).
Directionality: Water surfaces show strong directional energy concentration along horizontal and vertical axes due to wave textures, unlike the isotropic distribution in generic scenes.

2. Band-Selective and Spatially Routed (BSSR) Module

To address the spectral findings, the authors inserted the BSSR module exclusively into the P2 high-resolution branch (stride 4) of the detector. This ensures enhancement occurs where small targets are spatially visible without overloading the entire network. The module consists of:

Frequency Branch: Decomposes features into Low, Mid, and High-frequency bands. Unlike standard approaches that suppress mid-frequencies as noise, BSSR learns to route and emphasize the mid-frequency band, which contains substantial target cues in water scenes.
Directional Branch: Applies four-directional Sobel kernels (horizontal, vertical, diagonal) to capture fragmented edge responses caused by wave interference.
Local Context Branch: Uses a 5×5 depthwise convolution to preserve local spatial support.
Adaptive Gating: A lightweight router generates sigmoid-activated gates ( $\alpha_1, \alpha_2, \alpha_3$ ) to fuse the enhanced features with the original input. The output formula is:
$F_{out} = F + \alpha_1 F_{base} + \alpha_2 F - \alpha_3 |F - F_{base}|$
This structure injects enhanced cues, preserves the original signal, and adaptively suppresses residual discrepancies (acting as a denoising gate).

3. Scale-Normalized Kalman IoU (SNK-IoU) Loss

The authors introduce a training-only loss function that adapts bounding-box regression supervision based on prediction quality, inspired by the Kalman filter gain mechanism.

Mechanism: It treats the IoU error ( $P = 1 - \text{IoU}$ ) as "prediction uncertainty" and the aspect-ratio deviation ( $R$ ) as "observation noise."
Adaptive Gain: A bounded Kalman gain ( $K \in [0.2, 0.8]$ ) is computed to mix linear and squared CIoU terms:
$L_{SNK} = K \cdot L_{CIoU} + (1-K) \cdot L^2_{CIoU}$
Behavior: When IoU is low (poor overlap), $K$ is high, favoring the linear term for strong correction. When IoU is high (good overlap), $K$ is low, favoring the squared term for fine-grained refinement. This adds zero computational cost during inference.

Key Contributions

Spectral Characterization: A cross-dataset analysis establishing that water-surface detection is a weak-target recognition problem characterized by low-SNR, mid-frequency-dominant backgrounds.
BSSR Module: A reusable feature-routing design that performs frequency decomposition, directional edge routing, and adaptive gating specifically within the high-resolution small-target branch.
SNK-IoU Loss: A sample-adaptive regression loss that balances coarse correction and fine refinement using a bounded gain analogy, improving localization without altering the inference graph.
Comprehensive Evaluation: Extensive experiments on the FloW benchmark and cross-dataset transfer tests on PoTATO, evaluating accuracy, small-object AP, model size, latency, and throughput.

Experimental Results

The proposed method was evaluated on the FloW dataset (floating waste detection) and tested for transferability on PoTATO.

Accuracy: On FloW, the proposed method achieved 47.35% mAP50:95, a 5.91 percentage point improvement over the YOLO11n baseline (41.44%).
Small Object Performance: The method improved AP(S) (small objects) by 3.30 points (from 30.50% to 33.80%) under COCO-standard metrics.
Efficiency:
- Parameters: Increased only by 0.121 M (total 2.71 M).
- Latency: On an NVIDIA RTX 5090, the model runs at 93.1 FPS (10.7 ms/frame).
- Deployment: When exported to ONNX and run on an RTX 4080 Super, the end-to-end latency is 6.40 ms (156.2 FPS).
Generalization: The BSSR module improved performance across four different lightweight detector families (YOLOv8n, YOLO11n, YOLO12n, YOLO26n) with gains ranging from 2.23 to 5.91 mAP50:95. Cross-dataset transfer (FloW $\to$ PoTATO and vice versa) showed consistent improvements (~4.5–4.7 points), suggesting the method captures general water-surface properties rather than dataset-specific patterns.
Convergence: The model converged faster (epoch 81) compared to other loss variants (e.g., Inner-IoU at epoch 188).

Significance and Claims

The paper claims that frequency-informed feature routing and inference-neutral adaptive supervision are effective strategies for improving weak-target detection in low-SNR dynamic backgrounds.

Design Principle: The work demonstrates that detector design should be tied to measured scene statistics (specifically frequency decay and directional energy) rather than relying solely on generic attention mechanisms or global scaling.
Real-Time Viability: The approach successfully balances accuracy and latency, making it suitable for real-time USV perception where computational resources are constrained.
Modesty and Limitations: The authors explicitly state that while the model exceeds real-time requirements on desktop GPUs, embedded deployment (e.g., on NVIDIA Jetson Orin) and TensorRT optimization remain future work. They also acknowledge limitations regarding:
- Targets smaller than ~6×6 pixels (below the P2 stride-4 resolution limit).
- False positives triggered by strong specular reflections that mimic target edges.
- The lack of a held-out test set in the FloW benchmark (though mitigated by cross-dataset testing).
- The inability of visible-light imagery to handle night, fog, or heavy rain without additional sensors.

The authors position this work as a decision-assistance tool for USVs, emphasizing that field deployment requires temporal consistency checks and human supervision, rather than claiming autonomous safety certification.